VoxAgent: a multi-tenant voice agent platform
A SaaS where each tenant builds its own Gemini Live voice agent, connects it to a phone number, gives it tools and feeds it a knowledge base.
- Client
- Kentron Technologies (internal product)
- Sector
- AI infrastructure
- Shipped
- 2026
- Practice
- AI products and agents
VoxAgent is a multi-tenant platform for building and running voice agents. A tenant signs up, brings its own Gemini API key and its own phone number, designs an agent in the browser, and talks to it over a WebSocket before putting it on a real telephone line through Tata Smartflo. The agent can call functions in the tenant's own systems and answer from documents the tenant has uploaded. Every tenant sees only its own data.
The problem
Voice agent demos are easy and voice agents in production are not. The hard parts are the ones nobody shows in a demo: streaming audio in both directions without a half-second of dead air, a phone call that behaves the same as the browser test, and strict separation so one tenant's API key, call recordings and documents can never be read by another.
What we built
- 01
A browser client that streams microphone audio over a WebSocket to Gemini Live and plays the reply as it arrives, so the agent can be interrupted mid-sentence.
- 02
Bring-your-own-key tenancy: each tenant stores its own Gemini credentials and its own phone number, so usage and cost sit with the tenant rather than being pooled.
- 03
Telephony through Tata Smartflo, so the same agent that was tested in a browser tab answers an actual inbound call.
- 04
Function calling, where a tenant declares the tools its agent may invoke and the platform executes them against the tenant's own endpoints.
- 05
A retrieval layer: documents are uploaded, parsed (including PDFs), chunked and embedded, so the agent answers from the tenant's material instead of improvising.
- 06
Row-level tenant scoping in Prisma, enforced on every query rather than filtered in the UI.
What it does now
- A tenant can go from signup to a working browser agent without talking to us.
- The same agent definition serves a browser test and a live phone call.
- Tool calls let an agent read and write in the tenant's existing systems during a conversation.
- Transcripts are stored per conversation, so an answer can be audited after the fact.
What people ask about this work.
Can a voice agent be connected to an Indian phone number?
Yes. We connected agents to Indian numbers through Tata Smartflo, so an inbound caller reaches the agent on an ordinary telephone line. The same agent definition that was tested in the browser handles the call.
Who pays for the AI usage in a multi-tenant voice platform?
In VoxAgent each tenant brings its own Gemini API key and its own phone number, so model usage and telephony are billed to the tenant directly. This keeps cost attribution simple and means one tenant's traffic cannot exhaust another's quota.
How do you stop a voice agent from inventing answers?
By giving it a retrieval layer. Documents are uploaded, parsed, chunked and embedded, and the agent answers from that material. Tool calls fetch live data from the tenant's own systems rather than being recalled from the model's memory, and every conversation keeps a transcript so an answer can be checked afterwards.